Intelligent device fault prediction method and system based on ai multi-modal large model

Abstract

The application relates to the technical field of intelligent equipment fault diagnosis, and discloses an intelligent equipment fault prediction method and system based on an AI multi-modal large model, the method comprising the following steps: classifying and performing heterogeneous signal space-time alignment on multi-source data streams; calculating sensor drift bias distribution and dynamically correcting signals by using a recursive least square algorithm; separating background noise and reconstructing and enhancing weak fluctuation components; performing unified feature space mapping and multi-level integration; extracting abnormal features based on time series modeling and dynamic confidence intervals; constructing a multi-dimensional state trajectory, analyzing evolution trends and change rates by using a long short-term memory network, and calculating fault risk probabilities.The application can effectively eliminate heterogeneous data differences and environmental noise interference, and realizes dynamic tracking and accurate early warning of early weak fault risks of equipment.

Description

Technical Field

[0001] This invention relates to the field of intelligent device fault diagnosis technology, and in particular to an intelligent device fault prediction method and system based on an AI multimodal large model. Background Technology

[0002] With the widespread application of IoT chips, high-precision sensors, and edge computing technologies, the data generated during equipment operation is experiencing explosive growth. This data covers multiple modes, including vibration and temperature, and contains crucial information about the health status of the equipment. How to utilize this massive amount of multi-source data to detect potential risks before serious equipment failures occur has become a critical requirement for ensuring the continuous operation of systems.

[0003] In existing technologies, fault prediction primarily relies on univariate threshold monitoring or independent signal analysis methods based on traditional statistics. The system directly collects raw data from various sensors (such as motor vibration frequency or pipeline pressure values), performs simple data cleaning according to their independent signal channels, and compares it with pre-set fixed safety thresholds or baselines. When the monitored value amplitude exceeds a specific limit of the physical quantity (e.g., significant severe vibration or a sudden temperature rise), the system determines that the equipment is abnormal and issues an alarm signal. However, this method struggles to uniformly understand and fuse multi-source heterogeneous data under complex operating conditions within the same feature space. This results in weak early fault characteristics hidden in massive fluctuations of normal operating signals, environmental noise, and sensor drift being easily masked by vastly different background signals. Traditional methods cannot effectively distinguish between normal fluctuations and early weak faults, often only detecting significant faults in later stages of development, and are prone to numerous false alarms or missed alarms in early warning.

[0004] In summary, existing technologies have the problem of accurately identifying subtle early fault characteristics in complex environments where multiple heterogeneous data coexist. Summary of the Invention

[0005] This invention provides a method and system for predicting faults in intelligent devices based on an AI multimodal large model, in order to solve the problem of accurately identifying weak early fault characteristics in complex environments where multi-source heterogeneous data coexist.

[0006] Firstly, to address the aforementioned technical problems, this invention provides a method for predicting faults in intelligent devices based on an AI multimodal large model, comprising: Acquire multi-source data streams during the operation of intelligent devices, classify and group the multi-source data streams and align them with heterogeneous signals in time and space to obtain structured multi-source signal data; Based on the structured multi-source signal data, sensor drift characteristic parameters are extracted and deviation distribution is calculated to obtain the deviation distribution. Based on the deviation distribution, the recursive least squares algorithm is used to perform real-time dynamic correction of the signal deviation to obtain the calibrated steady-state signal sequence. Based on the calibrated steady-state signal sequence, background noise separation and weak fluctuation components are extracted, and the weak fluctuation components are reconstructed and enhanced to obtain the enhanced feature signal. Based on the enhanced feature signal, dimensionless transformation and unified feature space mapping are performed to obtain the mapped feature vector. Then, multi-level hierarchical integration is performed on the mapped feature vector to obtain the fused comprehensive feature representation. Based on the fused comprehensive feature representation, time-series evolution law modeling and dynamic prediction deviation analysis are performed to obtain an abnormal feature set; Based on the set of abnormal features, a multidimensional state trajectory is constructed. The evolution trend of the multidimensional state trajectory and the probability of failure risk are calculated using a long short-term memory network to obtain the failure prediction result of the intelligent device.

[0007] Secondly, the present invention provides an intelligent device fault prediction system based on an AI multimodal large model, comprising: The data processing module is used to acquire multi-source data streams during the operation of intelligent devices, classify and group the multi-source data streams and align them with heterogeneous signals in time and space to obtain structured multi-source signal data. The drift calibration module is used to extract sensor drift characteristic parameters and calculate deviation distribution based on the structured multi-source signal data, obtain the deviation distribution, and use the recursive least squares algorithm to perform real-time dynamic correction of signal deviation based on the deviation distribution to obtain the calibrated steady-state signal sequence. The signal enhancement module is used to separate background noise and extract weak fluctuation components based on the calibrated steady-state signal sequence, and to reconstruct and enhance the weak fluctuation components to obtain the enhanced feature signal. The feature fusion module is used to perform dimensionless transformation and unified feature space mapping based on the enhanced feature signal to obtain the mapped feature vector, and to perform multi-level hierarchical integration on the mapped feature vector to obtain the fused comprehensive feature representation. An anomaly detection module is used to perform time-series evolution law modeling and dynamic prediction deviation analysis based on the fused comprehensive feature representation to obtain an anomaly feature set; The risk prediction module is used to construct a multi-dimensional state trajectory based on the set of abnormal features, and to perform evolution trend analysis and fault risk probability calculation on the multi-dimensional state trajectory using a long short-term memory network to obtain the fault prediction result of the intelligent device.

[0008] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention preprocesses heterogeneous information representation by adopting a standardized method and converts various signals into a unified feature space through feature mapping technology. This method can solve the natural isolation and differences in physical characteristics, acquisition frequency and expression form of the original multi-source data. At the same time, it can eliminate the problems of time asynchrony and unit inconsistency between heterogeneous data, thereby realizing the unified understanding and comprehensive utilization of multi-source heterogeneous data, making model training more stable, significantly improving the accuracy of anomaly identification, and reducing the computing resource occupation and memory consumption of edge devices.

[0009] (2) This invention removes background noise interference by combining signal enhancement technology and uses dynamic calibration algorithms (such as recursive least squares algorithm) to correct deviations caused by sensor drift in real time. This method can separate weak early fault signals hidden in normal fluctuation confusion in strong noise environment and automatically cancel the data drift caused by equipment aging. This effectively avoids false alarms and missed alarms caused by environmental noise or sensor instability, improves the robustness of the system in long-term operation and the sensitivity to capture weak fault characteristics, and does not require frequent manual calibration, thus reducing maintenance costs.

[0010] (3) This invention integrates signals from different sources in a hierarchical manner by constructing a multi-layer feature fusion framework, and applies an anomaly detection model with a dynamic confidence interval to perform real-time analysis of the integrated features after fusion. This method realizes deep correlation and accurate analysis of multi-dimensional spatiotemporal features, and can adaptively adjust the detection threshold according to environmental changes (such as seasons and time periods), realize continuous tracking of the dynamic trend of equipment status, accurately distinguish between normal fluctuations and potential anomalies, thereby improving the timeliness and reliability of early fault risk prediction, and providing efficient technical support for equipment health management. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the intelligent device fault prediction method based on AI multimodal large model provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of an intelligent device fault prediction system based on an AI multimodal large model provided in the second embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] Reference Figure 1 The first embodiment of the present invention provides a method for predicting faults in intelligent devices based on an AI multimodal large model, comprising the following steps: S11, acquire multi-source data streams during the operation of the smart device, classify and group the multi-source data streams and align them with heterogeneous signals in time and space to obtain structured multi-source signal data; S12, based on the structured multi-source signal data, extract sensor drift characteristic parameters and calculate deviation distribution to obtain deviation distribution, and use the recursive least squares algorithm to perform real-time dynamic correction of signal deviation based on the deviation distribution to obtain a calibrated steady-state signal sequence. S13, Based on the calibrated steady-state signal sequence, perform background noise separation and extract weak fluctuation components, and reconstruct and enhance the weak fluctuation components to obtain the enhanced feature signal; S14. Based on the enhanced feature signal, dimensionless transformation and unified feature space mapping are performed to obtain the mapped feature vector. Then, multi-level hierarchical integration is performed on the mapped feature vector to obtain the fused comprehensive feature representation. S15, Based on the fused comprehensive feature representation, perform time-series evolution law modeling and dynamic prediction deviation analysis to obtain an abnormal feature set; S16. Based on the set of abnormal features, construct a multidimensional state trajectory, and use a long short-term memory network to perform evolution trend analysis and fault risk probability calculation on the multidimensional state trajectory to obtain the fault prediction result of the intelligent device.

[0014] In step S11, multi-source data streams during the operation of the intelligent device are acquired, and the multi-source data streams are classified, grouped, and spatiotemporally aligned with heterogeneous signals to obtain structured multi-source signal data, including: Acquire multi-source data streams from various sensors; Based on the multi-source data stream, physical characteristic identification is parsed and acquisition frequency values ​​are extracted to obtain the original signal sequence; Based on the original signal sequence, channel mapping based on preset physical characteristic identifiers is performed to obtain heterogeneous signal groups containing high-frequency and low-frequency signals. Based on the heterogeneous signal grouping, downsampling interpolation is performed for the high-frequency signal and forward hold timing alignment is performed for the low-frequency signal to obtain the aligned signal sequence; Based on the aligned signal sequence, time-domain distribution characteristics are calculated and structured model mapping is performed to obtain structured multi-source signal data.

[0015] First, the data acquisition module reads multi-source data streams in real time from sensor interfaces distributed at key monitoring points of the smart device via industrial fieldbuses (such as Modbus TCP and Profinet) or IoT transmission protocols (such as MQTT). After receiving the data packets, the system parses the metadata in the data frame header, extracts the physical interface address code and the current acquisition frequency value, and immediately uses a local clock synchronized with Network Time Protocol (NTP) to stamp each frame of data with a unified millisecond-level system timestamp, generating a raw signal sequence with timestamps.

[0016] Next, the system performs a channel mapping operation based on physical characteristic identifiers. During this process, the system calls a device mapping table pre-stored in non-volatile memory. This mapping table is a database of relationships generated during the system initialization configuration phase based on the device electrical schematic and sensor layout diagram. It records a one-to-one correspondence between each physical interface address code (Key), logical channel identifier (Value), and physical quantity type (such as "stator temperature" or "bearing vibration"). Based on the physical interface address code parsed from the original signal sequence, the system performs a hash lookup match in the mapping table, routes the original signal to the corresponding logical processing channel, and groups the signals according to the physical quantity type and acquisition frequency value defined in the mapping table: vibration, current, and other signals with acquisition frequencies higher than a preset frequency threshold (e.g., 100Hz) are classified as high-frequency signals; temperature, pressure, and other signals with acquisition frequencies lower than or equal to the preset frequency threshold are classified as low-frequency signals, thus obtaining heterogeneous signal groups containing both high-frequency and low-frequency signals.

[0017] It should be noted that the construction rules for the aforementioned device mapping table should follow the address space model specifications in the OPC UA unified architecture standard to ensure that the physical addresses of devices from different manufacturers can be uniformly mapped to standardized logical node IDs, achieving hardware independence. The setting of the preset frequency threshold (e.g., 100Hz) is based on the frequency band division recommendations for vibration signals (usually requiring kHz-level sampling) and process parameters (such as temperature and flow rate, usually Hz-level or lower) in the GB / T 19873.1 "Machine Condition Monitoring and Diagnosis" standard; those skilled in the art can determine the specific cutoff frequency based on the dynamic response characteristics of the specific monitored object (e.g., motor rotor time constant) and the sensor technical specifications.

[0018] Subsequently, the system constructs a unified time grid with a fixed time step (e.g., 1 second) and performs time alignment operations on heterogeneous signal groups. For high-frequency signal groups, the system performs downsampling interpolation, that is, within each time grid window, it extracts all sampling points within the corresponding time period, calculates their effective value (RMS) or arithmetic mean, and uses this single value as the representative value of that grid moment. For low-frequency signal groups, the system performs forward hold, that is, when no new data arrives at the current grid moment, it automatically reuses the most recent valid sampling value from the previous moment to fill the current grid, until the next frame of new data arrives, thereby uniformly mapping all signals of different frequencies to the same time axis coordinates to obtain the aligned signal sequence. Next, the system calculates the temporal distribution characteristics of the aligned signal sequence based on a feature calculation window (e.g., a sliding window length of 60 seconds and a step size of 1 second). The window length is set based on the assumption of signal stationarity and an extension of the Nyquist sampling theorem. Specifically, the window length should cover at least 3 to 5 typical fluctuation cycles of the monitored physical quantity (especially low-frequency process quantities), or refer to the recommended duration for steady-state measurement of machine vibration in ISO 10816 standard, to ensure that the calculated statistical characteristics are representative and stable, and can filter out transient random interference. Within each window, the system calculates the mean, variance, and peak-to-peak value of the data from each channel to construct a feature vector.

[0019] Finally, the system performs a structured model mapping operation to construct the device state matrix. This matrix is ​​a multidimensional tensor, typically defined as T×N×F, where T represents the length of the time step sequence (corresponding to the historical observation window), N represents the total number of sensor logic channels, and F represents the number of feature dimensions calculated for each channel (such as mean, variance, etc.). The system fills the corresponding positions of the tensor with the calculated time-domain distribution feature values ​​of each channel, strictly according to the time order and channel index. In this way, the originally discrete and heterogeneous time series is transformed into structured multi-source signal data that can be directly processed by a computer, preparing it for formatted input to subsequent deep learning models.

[0020] In step S12, based on the structured multi-source signal data, sensor drift characteristic parameters are extracted and deviation distribution is calculated to obtain the deviation distribution. Then, based on the deviation distribution, a recursive least squares algorithm is used to dynamically correct the signal deviation in real time, resulting in a calibrated steady-state signal sequence, including: Based on the structured multi-source signal data, segmented statistical analysis is performed to obtain the signal amplitude within the current time window. The signal amplitude within the current time window is then compared with the pre-stored historical stable period reference amplitude to obtain the zero-point offset and sensitivity attenuation. Based on the zero-point offset and the sensitivity attenuation, frequency band mapping is performed to obtain the deviation distribution; Based on the deviation distribution, iterative calculations based on the recursive least squares algorithm are performed until the preset error tolerance range is reached to obtain dynamic compensation parameters; Based on the dynamic compensation parameters and the structured multi-source signal data, linear weighting and multiplication correction are performed to obtain the calibrated steady-state signal sequence.

[0021] First, the drift calibration module receives the structured multi-source signal data output in step S11 and performs segmented statistical analysis. The system defines a sliding observation window with configurable length. Within each window, the arithmetic mean of the signal is calculated as the current DC component, and the standard deviation of the signal is calculated as the current AC energy. Then, the system calls the pre-stored historical stable period reference amplitude stored in the database. This reference amplitude includes the reference mean and reference standard deviation of the device in its initial healthy state. The system performs a comparison calculation, subtracting the reference mean from the arithmetic mean of the current window. The difference is defined as the zero-point offset, which characterizes the degree of drift of the sensor's static operating point. Simultaneously, the standard deviation of the current window is divided by the reference standard deviation, and the quotient is defined as the sensitivity attenuation, which characterizes the gain change of the sensor's dynamic response capability. Next, the system maps the zero-point offset to a low-frequency drift characteristic and the sensitivity attenuation to a high-frequency gain characteristic based on physical characteristics, combining them to obtain the deviation distribution.

[0022] Subsequently, the system initiates an iterative calculation process based on the recursive least squares algorithm to solve for the dynamic compensation parameters. This process approximates the optimal solution through pure numerical iteration. At each time step, the system first calculates the gain vector. Specifically, it performs matrix multiplication on the covariance matrix updated at the previous time step and the input feature vector at the current time step to obtain an intermediate vector. Simultaneously, it calculates the weighted quadratic form of the current input feature vector with respect to the covariance matrix and adds this value to the forgetting factor (usually set to 0.95 to 0.99) to obtain a scalar denominator. Finally, it divides the intermediate vector by the scalar denominator to obtain the gain vector at the current time step. This vector reflects the correction weight of the parameter update by the new data.

[0023] Next, the system uses the estimated parameters from the previous time step to predict the current output. Specifically, it extracts the low-frequency and high-frequency compensation coefficients obtained from the previous iteration, constructs them into a parameter estimation vector, and performs a vector dot product operation (i.e., summing corresponding elements) with the current input feature vector to generate the theoretical prediction value for the current time step. Subsequently, the system subtracts the theoretical prediction value from the currently observed deviation distribution value to obtain the prediction error. Then, the system multiplies the gain vector with the prediction error and adds the product to the parameter estimation vector from the previous time step, thereby updating the dynamic compensation parameters for the current time step. This iterative process continues until the absolute value of the prediction error is less than a preset error tolerance range, at which point the finally converged dynamic compensation parameters are output. Finally, the system uses the dynamic compensation parameters to linearly correct the original signal. Specifically, it subtracts the calculated low-frequency compensation coefficient (i.e., zero-point offset) from the original signal value, and then divides the result by the high-frequency compensation coefficient (i.e., sensitivity gain), thereby eliminating the effects of drift and attenuation to obtain the calibrated steady-state signal sequence.

[0024] It should be noted that the setting of the pre-stored historical stability period benchmark range is based on the "fingerprint extraction" principle in the equipment's full life cycle management. Typically, after the equipment has been installed, debugged, and run for 24 hours, and all indicators are confirmed to be in optimal condition, data is continuously collected for a period of time, and its statistical characteristic values ​​are calculated and solidified, serving as the reference point for all subsequent health assessments. The preset error tolerance range is set primarily based on the linearity error index or resolution limit specified in the sensor datasheet. A common setting method is to set this tolerance between one ten-thousandth and five thousandths of the sensor's full-scale output value. For example, for a sensor with a range of 0 to 10V and an accuracy of 0.1%, the error tolerance can be set to 0.01V, ensuring calibration accuracy while preventing the algorithm from falling into an infinite loop due to excessive pursuit of minimal errors. The forgetting factor (usually between 0.95 and 0.99) is set to balance the algorithm's tracking speed and stability. The closer this value is to 1, the longer the algorithm remembers, the less sensitive it is to noise, but the slower it tracks sudden changes. In this embodiment, it is set empirically based on the sensor's expected maximum drift rate.

[0025] In step S13, based on the calibrated steady-state signal sequence, background noise separation and weak fluctuation components are extracted, and the weak fluctuation components are reconstructed and enhanced to obtain enhanced feature signals, including: Based on the calibrated steady-state signal sequence, the local signal-to-noise ratio is calculated, and signal segments with local signal-to-noise ratios lower than a preset weak feature screening threshold are selected to obtain a weak feature subset. Based on the weak feature subset, wavelet packet decomposition and soft threshold shrinkage denoising are performed to obtain a denoised feature vector sequence. Based on the denoised feature vector sequence, Butterworth low-pass filtering and residual Hilbert transform analysis are performed to obtain the early anomaly hidden signal; Based on the early anomaly hidden signal, sinusoidal superposition reconstruction and reference waveform superposition are performed to obtain the enhanced feature signal; The method of obtaining a weak feature subset further includes: For vibration or acoustic signature-type high-frequency non-stationary signals in the calibrated steady-state signal sequence, calculate the kurtosis or spectral kurtosis of the corresponding signal segment; Signal segments with kurtosis or spectral kurtosis values ​​higher than a preset kurtosis threshold are selected as part of the weak feature subset.

[0026] First, the signal enhancement module receives the calibrated steady-state signal sequence output from step S12 and performs background noise assessment. The system defines a fixed-length assessment window, and calculates the ratio of signal power to noise power within each window. Specifically, the mean square of the signal amplitude within the window is calculated as the signal power, and the variance of the signal amplitude minus the mean is calculated as the noise power. The signal power is then divided by the noise power to obtain the local signal-to-noise ratio (SNR). Next, the system compares the calculated local SNR with a preset weak feature screening threshold. The system logic is set so that when the local SNR is lower than the threshold, the segment is determined to contain early weak fault features masked by background noise. The system extracts and marks these specific signal segments, compiling them into a weak feature subset. When the local SNR is higher than or equal to the threshold, the segment is determined to be a clear signal or a strong fault signal. The system directly retains the calibrated steady-state signal sequence within this time period without performing subsequent reconstruction and enhancement operations to maintain the original authenticity of the signal. The preset weak feature screening threshold is set based on the background noise statistical distribution of the equipment under normal operating conditions. It is usually set to 30% to 50% of the average signal-to-noise ratio of the equipment under no-load or healthy conditions. For example, if the average signal-to-noise ratio under healthy conditions is 20dB, the threshold can be set to 6dB to 10dB. The purpose is to specifically capture those early fault signals with weak energy that are easily drowned out by environmental noise (such as the impact caused by slight peeling of bearing balls) and prevent them from being ignored in the conventional averaging process.

[0027] It should be noted that the local signal-to-noise ratio (SNR) calculation method based on power ratio described in this embodiment is mainly applicable to stationary process signals such as temperature, pressure, and liquid level, which are dominated by DC trends. In this case, signal fluctuations are usually considered as measurement noise. However, for high-frequency non-stationary signals such as vibration and acoustic signatures, which are dominated by AC components, their fault characteristics often manifest as fluctuations in signal energy (i.e., increased variance). Directly using variance as the denominator to calculate the SNR may lead to the suppression of fault characteristics. Therefore, in the preferred embodiment of this invention, for such high-frequency non-stationary signals, the system uses "kurtosis" or "spectral kurtosis" as the screening criterion for weak features. Specifically, the system calculates the ratio of the fourth central moment of the signal segment to the fourth power of the standard deviation. This index is extremely sensitive to the impact components in the signal. For background random noise that follows a Gaussian distribution, its kurtosis value is approximately 3. When the calculated kurtosis value is significantly greater than 3 (e.g., exceeding the preset kurtosis threshold of 3.5), it indicates that the signal segment contains non-Gaussian transient impact components (such as periodic pulses caused by early pitting corrosion of bearings). The system then determines the segment as a subset containing weak features for subsequent processing, thereby avoiding the limitations of conventional signal-to-noise ratio calculation in vibration analysis.

[0028] It should be further explained that the calculation and screening of local signal-to-noise ratio described in this invention is a general logical framework for feature screening. In practical applications, the calculation and judgment of local signal-to-noise ratio can be specified into different mathematical indicators for signals with different physical characteristics. For stationary signals such as temperature and pressure, signal-to-noise ratio (SNR) based on power ratio is effective. For non-stationary signals such as vibration and acoustic emission, their early faults often manifest as transient impacts, making it more suitable to use statistical indicators sensitive to pulse components, such as kurtosis or spectral kurtosis, for equivalent screening. Therefore, the preset weak feature screening threshold is specifically expressed as a signal-to-noise ratio threshold or a kurtosis threshold depending on the signal type. Those skilled in the art can select and set the corresponding indicators and thresholds during the system configuration stage based on prior knowledge of the sensor type and the monitored object.

[0029] Subsequently, the system performs wavelet packet decomposition on the weak feature subset to obtain a series of wavelet packet coefficients in different frequency bands, and calculates a preset denoising threshold. The specific calculation process is as follows: first, the first layer of high-frequency detail coefficients obtained by decomposition is extracted, the median of the absolute value of the detail coefficient is calculated, and the median is divided by a constant 0.6745 to obtain the noise standard deviation estimate. Then, the square root of twice the natural logarithm of the signal length is calculated. Finally, the noise standard deviation estimate is multiplied by the square root to obtain a general adaptive denoising threshold. The system uses this threshold to perform soft threshold shrinkage denoising on the detail coefficients in the high-frequency part, and obtains the denoised feature vector sequence after inverse wavelet transform.

[0030] Next, the system performs reference separation on the denoised feature vector sequence, using a Butterworth low-pass filter to extract the low-frequency trend term in the signal as the reference signal waveform. Then, the reference signal waveform is subtracted from the denoised feature vector sequence to obtain a residual sequence containing only high-frequency transient impulse components. The system performs a Hilbert transform on the residual sequence. Specifically, the residual sequence is first convolved with the time-domain impulse response function to generate an imaginary part signal with a 90-degree phase lag. Then, the original residual sequence is used as the real part, and the imaginary part signal is used as the imaginary part to construct an analytic signal in the complex domain. Next, the magnitude of the analytic signal at each time step (i.e., the square root of the sum of the squares of the real and imaginary parts) is calculated to obtain the instantaneous amplitude envelope. Finally, the time derivative of the arctangent function of the ratio of the imaginary to the real part of the analytic signal is calculated to obtain the instantaneous frequency. Finally, the system performs a signal reconstruction operation. Specifically, the extracted instantaneous amplitude envelope is used as the time-varying amplitude coefficient, and the integral of the instantaneous frequency over time is used as the phase parameter. These parameters are substituted into the standard sine function formula to generate a pure oscillating waveform. This pure oscillating waveform is then numerically superimposed on the previously separated reference signal waveform at corresponding time points to synthesize an enhanced weak feature segment. Furthermore, the system splices and fuses this enhanced weak feature segment with the previously retained conventional signal segment according to the original timestamps, forming a continuous and complete enhanced feature signal on the time axis.

[0031] It should be noted that the common method for setting the preset noise threshold is to estimate the background noise level by statistically analyzing the fluctuation of the highest frequency detail component in the signal. This setting can be dynamically adjusted adaptively with changes in signal length and noise intensity, effectively eliminating Gaussian white noise while preserving signal spikes containing fault information to the greatest extent. The Hilbert transform and analytical signal construction process described above is the standard mathematical paradigm for demodulating amplitude-frequency modulation (AM-FM) signals in current industrial signal processing. Its physical meaning lies in decomposing non-stationary vibration signals into instantaneous energy changes (amplitude) and frequency changes (frequency), thereby accurately capturing the microsecond-level impact characteristics generated by bearings or gears in the early stages of a fault.

[0032] In step S14, based on the enhanced feature signal, dimensionless transformation and unified feature space mapping are performed to obtain the mapped feature vector. Then, multi-level hierarchical integration is performed on the mapped feature vector to obtain the fused comprehensive feature representation, including: Based on the enhanced feature signal, statistical distribution moments are calculated and dimensionless operations are performed to obtain a normalized sequence. Based on the normalized sequence, principal component analysis projection and L1 norm sparsity constraint are performed to obtain the mapped feature vector. Based on the mapped feature vector, frequency band hierarchy is divided and time delay deviation is calculated to obtain the association weight of each level; Based on the association weights and the mapped feature vectors, a weighted concatenation and integration are performed to obtain the fused comprehensive feature representation.

[0033] First, the feature fusion module receives the enhanced feature signal output in step S13 and performs statistical distribution moment calculation and dimensionless operation on it. The system calculates the arithmetic mean and standard deviation of the signal in the current data window. Then, using the Z-score normalization method, the arithmetic mean is subtracted from each data point in the signal, and the difference is divided by the standard deviation to obtain a normalized sequence with a mean of 0 and a variance of 1.

[0034] Next, the system performs principal component analysis projection on the normalized sequence. This process first requires constructing a covariance matrix. Specifically, the matrix formed by the normalized sequence is transposed to obtain a transpose matrix. This transpose matrix is ​​then multiplied with the original normalized sequence matrix. Each element of the resulting product matrix is ​​divided by the sample size minus one, thus obtaining a square matrix reflecting the linear correlation between features. Subsequently, the system performs eigenvalue decomposition on the covariance matrix. Specifically, this involves constructing the characteristic equation of the covariance matrix (i.e., the equation with a determinant equal to zero) and solving this equation to obtain... The series of scalar roots obtained are the eigenvalues. Each eigenvalue is substituted into the linear equation system to solve for the corresponding non-zero solution vector, thus obtaining the eigenvector. The system then calculates the cumulative contribution rate. Specifically, all eigenvalues ​​are sorted in descending order of numerical value, and the arithmetic sum of the first N eigenvalues ​​is calculated as the cumulative variance. The cumulative variance is divided by the sum of all eigenvalues ​​to obtain the cumulative contribution rate. The eigenvectors that make this ratio greater than the preset contribution rate threshold are selected to construct the projection matrix. The normalized sequence is projected to the low-dimensional space, and L1 norm sparsity constraints are applied to remove redundant features, thus obtaining the mapped eigenvector.

[0035] Subsequently, the system performs frequency band attribute labeling on the original feature sources before mapping or the normalized sequences before PCA projection. Specifically, based on the sensor type and the frequency domain characteristics in the signal processing process, the system divides the features into low-frequency level (corresponding to process quantity features such as temperature and pressure), mid-frequency level (corresponding to rotational speed and fundamental frequency vibration features), and high-frequency level (corresponding to early shock and harmonic features). For each level, the system calculates the time delay bias using a cross-correlation function. Specifically, the feature sequence of the current level is shifted relative to the historical baseline sequence along the time axis. At each shift step, the sum of the products of corresponding points in the two sequences is calculated to generate a cross-correlation curve. The moment of the curve's peak is identified, and the time difference between this peak moment and the zero-lag moment is calculated as the time delay bias. Finally, the system calculates the association weights for each level. The specific calculation logic involves taking the time delay bias as the negative exponent of the natural constant e and performing a power operation to obtain the initial weights. Then, the initial weight of a certain level is divided by the sum of the initial weights of all levels to obtain the normalized association weights. The system multiplies the feature vectors of each level by their corresponding association weights and concatenates them along the feature dimension to obtain the fused comprehensive feature representation.

[0036] It should be noted that the preset contribution rate threshold is usually set based on the Pareto principle (80 / 20 rule) in statistics and signal compression theory, and is generally set to 85% to 95%, meaning that the retained principal components can explain most of the information variation in the original data. In this embodiment, 95% is selected as the threshold to balance computational efficiency and information integrity. The introduction of the L1 norm sparsity constraint is to solve the "curse of dimensionality" problem commonly found in multi-source heterogeneous data fusion. Its technical essence is to automatically select the feature subset that contributes the most to fault prediction through compressed sensing principle. The preset frequency band cutoff frequency is based on GB / T 19873 and ISO The 10816 standard defines, for example, frequencies below 10Hz as low frequency, frequencies from 10Hz to 1kHz as mid frequency, and frequencies above 1kHz as high frequency, to conform to the frequency domain distribution of different fault modes of physical equipment (such as imbalance, misalignment, and bearing damage). The physical meaning of the time delay deviation calculation is to evaluate the time asynchrony of multi-source signals during transmission and acquisition. By giving higher weight to low-delay signals, the model training noise caused by time alignment errors can be effectively reduced.

[0037] In step S15, based on the fused comprehensive feature representation, time-series evolution law modeling and dynamic prediction deviation analysis are performed to obtain an abnormal feature set, including: Based on the fused comprehensive feature representation, time-series evolution law learning and historical fluctuation variance analysis are performed to obtain the baseline prediction sequence and dynamic confidence interval; Based on the baseline prediction sequence and the fused comprehensive feature representation, the difference is calculated to obtain the residual sequence; Based on the residual sequence and the dynamic confidence interval, out-of-bounds point identification and likelihood calculation are performed to obtain potential outliers; Based on the potential anomalies, cluster center features are extracted to obtain an anomaly feature set.

[0038] First, the anomaly detection module receives the fused comprehensive feature representation output from step S14 and performs time-series evolution law learning and historical fluctuation variance analysis on it. The system uses a fixed-length sliding historical window to extract a segment of historical feature data and generates a baseline prediction sequence using a weighted moving average method. Specifically, the system assigns a time decay weight to each historical moment within the window. This weight decreases exponentially with the increase in time distance from the current moment (i.e., the closer the data is to the current moment, the greater the weight). The system multiplies the feature value of each moment within the window by its corresponding time decay weight and accumulates them. Then, the accumulated result is divided by the sum of all weights, and the resulting value is the trend prediction value for the current moment. The baseline prediction sequence is composed of the trend prediction values ​​of each moment in chronological order. At the same time, the system calculates the standard deviation of the feature data within the window and multiplies the standard deviation by a preset confidence coefficient to obtain the fluctuation tolerance. The baseline prediction sequence is then added to and subtracted from the fluctuation tolerance to determine the upper and lower limits of the dynamic confidence interval that changes dynamically over time.

[0039] Next, the system performs difference calculation on the baseline prediction sequence and the fused integrated feature representation, subtracting the value of the baseline prediction sequence at the corresponding time from the value of the integrated feature representation received in real time to obtain the residual sequence. The system determines whether the value in the residual sequence exceeds the range of the dynamic confidence interval. If it does, it is identified as an out-of-bounds point, and the likelihood value of the out-of-bounds point is calculated. The calculation process is based on pure mathematical operations using the probability density function of the Gaussian distribution. First, the difference between the out-of-bounds point value and the center value of the confidence interval is calculated, and this difference is divided by the window standard deviation to obtain the standardized deviation. Then, the square of the standardized deviation is calculated, and the square value is divided by 2 and the negative number is taken to obtain a negative exponent. Next, an exponential operation is performed with the natural constant e as the base and the negative exponent as the power. Finally, the exponential function value with the natural constant e as the base and the negative exponent as the power is directly calculated without multiplying by the normalization constant of the probability density function, thus obtaining the non-normalized likelihood strictly within the range (0,1]. This value reflects the probability that the current data point belongs to the normal distribution, and its value range is between (0,1]. The lower the value, the smaller the probability of belonging to the normal distribution, that is, the greater the probability of an anomaly. The system sets a preset likelihood judgment threshold (e.g., 0.01). When the calculated non-normalized likelihood is lower than this threshold, the point is finally judged as a potential outlier.

[0040] Finally, the system uses the DBSCAN clustering algorithm to extract cluster center features from the potential anomalies. First, the system iterates through all marked potential anomalies. For each point, it calculates the Euclidean distance between it and other adjacent potential anomalies on the time axis, and counts the number of adjacent points within a preset clustering radius. If this number exceeds a preset minimum density threshold, these points are grouped into the same anomaly cluster. Then, the system calculates the arithmetic mean of all points within the anomaly cluster across all feature dimensions, using the calculated mean vector as the geometric center of the anomaly cluster. This geometric center is the feature vector representing the anomaly event. Finally, the feature vectors of all cluster centers are aggregated to obtain the anomaly feature set. The preset clustering radius and minimum density threshold are key parameters of the DBSCAN clustering algorithm. A common method for setting the clustering radius is to calculate the average Euclidean distance between data points within the same fault event in historical fault cases, and set 1.2 to 1.5 times this average distance as the clustering radius to prevent misjudging the same fault fluctuation as multiple independent anomalies. The minimum density threshold is typically set to 3 to 5 points to ensure that occasional single-point noise is filtered out and only persistent fault characteristics are retained.

[0041] It should be noted that the preset confidence coefficient is set according to the 3σ principle in statistics, and is usually set to 3. This means that under the assumption of normal distribution, normal random fluctuations have a 99.7% probability of falling within this range, and any fluctuations outside this range are considered low-probability abnormal events. The preset likelihood threshold is set based on the tolerance of the system's false alarm rate, and is usually set to 0.05 or 0.01, corresponding to the significance level in the statistical significance test. In this embodiment, it is set to 0.01 to ensure that only high-confidence anomalies are captured.

[0042] In step S16, a multidimensional state trajectory is constructed based on the set of abnormal features. A long short-term memory network is then used to analyze the evolution trend and calculate the probability of failure risk on the multidimensional state trajectory to obtain the intelligent device failure prediction result, including: Based on the set of abnormal features, key parameters are extracted and multi-dimensional spatial mapping is performed to obtain state trajectory points; Based on the state trajectory points, a time-series evolution analysis is performed using a pre-trained long short-term memory network to obtain a dynamic trend curve; Based on the dynamic trend curve, the slope of the tangent is calculated to obtain the rate of change of the state trajectory points; Based on the rate of change of the state trajectory points, probabilistic inference is performed using a preset fault risk assessment model to obtain the fault prediction result of the intelligent device.

[0043] First, the risk prediction module receives the set of abnormal features output in step S15 and performs key parameter extraction and multi-dimensional space mapping operations on it. This process aims to reduce data redundancy and construct a visualized evolution space. Specifically, the system calculates the variance of each feature dimension in the abnormal feature set in the time dimension, sorts the variance values ​​in descending order, and selects the top K feature dimensions with a cumulative variance contribution rate of 90% as key parameters. Subsequently, the system uses these K key parameters as mutually orthogonal coordinate axes to define a K-dimensional Euclidean space, and projects the abnormal clustering center vector of each time step into this space to determine its unique spatial coordinate position. Then, these coordinate points are connected in chronological order to obtain a sequence of state trajectory points describing the device performance degradation path.

[0044] Next, the system inputs the state trajectory point sequence into a pre-trained Long Short-Term Memory (LSTM) network to perform temporal evolution analysis. The working mechanism of this network in the inference stage is that the input layer receives the current trajectory point coordinate sequence, the hidden layer uses the internal forget gate to discard redundant historical state information, uses the input gate to accept new information at the current moment and update the cell state, and uses the output gate to calculate the hidden state output at the current moment based on the updated cell state. After multiple recursive processing, the fully connected output layer maps the predicted state coordinate values ​​for several future time steps. The system connects these predicted values ​​into a smooth curve to obtain a dynamic trend curve. It should be noted that this network is built during the system's offline phase, and the training data comes from the operational data of similar devices throughout their entire historical lifecycle (i.e., a complete record from normal operation to complete failure). The network structure is typically configured to contain two hidden layers, each with 64 to 128 neurons to capture complex nonlinear degradation features. The training process adopts a supervised learning mode, using historical trajectory points as input and future real trajectory points as labels. The mean squared error (MSE) is selected as the loss function, and the Adam optimizer (with a learning rate set to 0.001) is used for iterative parameter updates until the prediction error on the validation set no longer decreases. Before being put into use, this network needs to undergo sufficient convergence verification to ensure that it has accurate time series extrapolation capabilities.

[0045] Subsequently, the system performs tangent slope calculation on the dynamic trend curve. Using the numerical differentiation method, it calculates the increment of the predicted point at the end of the curve relative to the predicted point at the previous moment, and divides the increment by the time step to obtain the tangent slope characterizing the degradation rate. The system defines the absolute value of this slope as the rate of change of the state trajectory point.

[0046] Finally, the system performs fault risk probability inference, inputting the calculated rate of change of the state trajectory points into a preset fault risk assessment model. This model is essentially a binary classification probability model, and its training data consists of historical maintenance records. The rate of change during normal equipment operation is marked as "0" (low risk), and the rate of change abruptly before a fault is marked as "1" (high risk). Parameter selection mainly involves regression weights and biases. The system uses maximum likelihood estimation (MLE) to fit and solve for these parameters to the historical data. The model's operation involves multiplying the input rate of change by the weight coefficients and adding a bias term to obtain a linear regression value. This linear regression value is then substituted into the Sigmoid activation function (i.e., 1 divided by 1 plus the negative power of the natural constant e) for nonlinear mapping, outputting a value between 0 and 1. The system multiplies this value by 100% to obtain the intelligent device fault prediction result, i.e., the percentage probability that the current device will experience a functional fault in the future. The bias term is typically negative (e.g., -5 to -10) and is used to adjust the activation center of the sigmoid function, ensuring that the output failure probability approaches 0 when the rate of change is zero or extremely low. The risk probability only rises rapidly to the high-risk range when the rate of change increases significantly and offsets the suppressive effect of the bias term.

[0047] Specifically, the training process of the fault risk assessment model can be achieved by collecting complete historical operating data of similar smart devices, including the entire process from a healthy state to a fault, defining a specific time period before the fault occurs as the warning period, typically 24-72 hours; for each time point in the historical data, the average rate of change of the state trajectory points within a previous time window (e.g., 1 hour) is calculated as the model input feature; if the time point falls within the warning period of a certain fault, the sample is labeled as "1" (high risk), otherwise it is labeled as "0" (low risk); using the constructed labeled model... The model is trained using a sample set with the goal of maximizing the log-likelihood function, i.e., minimizing the cross-entropy loss. Stochastic gradient descent (SGD) or the Adam optimizer is used for iterative parameter optimization. During training, the sample set is randomly divided into training, validation, and test sets in a 7:2:1 ratio. The validation set is used to monitor overfitting, and an L2 regularization term can be introduced. After training convergence, the obtained model parameters are fixed and the model is retrained periodically, such as quarterly or after accumulating a certain number of new fault cases. The new data is used to update the model parameters to achieve continuous model optimization.

[0048] It is worth noting that the AI ​​multimodal large model described in this invention refers to a deep learning framework capable of processing multi-source heterogeneous data. Its core lies in achieving a unified understanding of multimodal data through multi-level feature fusion and temporal modeling. In specific implementation, this large model can be built based on the Transformer architecture or an enhanced LSTM network. For example, a multi-head self-attention mechanism can be introduced in the feature fusion stage to perform cross-modal association learning of multi-source signals. The model training uses a large-scale historical dataset (covering multiple modalities such as vibration, temperature, and pressure). Parameters are optimized through pre-training and fine-tuning. The model size can be configured to a large network containing hundreds of millions of parameters according to actual needs to improve the generalization ability of feature extraction. Those skilled in the art can adapt the calibration and enhancement steps of this invention according to specific application scenarios and refer to the design principles of existing multimodal large models (such as Vision-Language Models) to ensure the accurate capture of early fault features.

[0049] In summary, this invention achieves dynamic tracking and early warning of early fault risks by performing structured classification and unified feature space mapping on multi-source heterogeneous signals, combining signal enhancement and sensor drift dynamic calibration techniques to extract weak features, and constructing a multi-layer feature fusion framework to apply an anomaly detection model to perform real-time analysis of comprehensive features.

[0050] Reference Figure 2 The second embodiment of the present invention provides an intelligent device fault prediction system based on an AI multimodal large model, comprising: The data processing module is used to acquire multi-source data streams during the operation of intelligent devices, classify and group the multi-source data streams and align them with heterogeneous signals in time and space to obtain structured multi-source signal data. The drift calibration module is used to extract sensor drift characteristic parameters and calculate deviation distribution based on the structured multi-source signal data, obtain the deviation distribution, and use the recursive least squares algorithm to perform real-time dynamic correction of signal deviation based on the deviation distribution to obtain the calibrated steady-state signal sequence. The signal enhancement module is used to separate background noise and extract weak fluctuation components based on the calibrated steady-state signal sequence, and to reconstruct and enhance the weak fluctuation components to obtain the enhanced feature signal. The feature fusion module is used to perform dimensionless transformation and unified feature space mapping based on the enhanced feature signal to obtain the mapped feature vector, and to perform multi-level hierarchical integration on the mapped feature vector to obtain the fused comprehensive feature representation. An anomaly detection module is used to perform time-series evolution law modeling and dynamic prediction deviation analysis based on the fused comprehensive feature representation to obtain an anomaly feature set; The risk prediction module is used to construct a multi-dimensional state trajectory based on the set of abnormal features, and to perform evolution trend analysis and fault risk probability calculation on the multi-dimensional state trajectory using a long short-term memory network to obtain the fault prediction result of the intelligent device.

[0051] It should be noted that the intelligent device fault prediction system based on AI multimodal large model provided in this embodiment of the invention is used to execute all the process steps of the intelligent device fault prediction method based on AI multimodal large model in the above embodiment. The working principle and beneficial effect of the two correspond one-to-one, so they will not be described again.

[0052] This invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a cold-plate intelligent processing control program. When the processor executes the computer program, it implements the steps described in the various embodiments of the cold-plate intelligent processing control method, for example... Figure 1 The step S11 shown. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the above system embodiments, such as the data processing module.

[0053] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0054] The electronic device may be a desktop computer, laptop, handheld computer, or smart tablet, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. It may include more or fewer components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0055] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting all parts of the electronic device via various interfaces and lines.

[0056] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0057] If the modules / units integrated into the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or system capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0058] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0059] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for predicting faults in intelligent devices based on an AI multimodal large model, characterized in that, include: Acquire multi-source data streams during the operation of intelligent devices, classify and group the multi-source data streams and align them with heterogeneous signals in time and space to obtain structured multi-source signal data; Based on the structured multi-source signal data, sensor drift characteristic parameters are extracted and deviation distribution is calculated to obtain the deviation distribution. Based on the deviation distribution, the recursive least squares algorithm is used to perform real-time dynamic correction of the signal deviation to obtain the calibrated steady-state signal sequence. Based on the calibrated steady-state signal sequence, background noise separation and weak fluctuation components are extracted, and the weak fluctuation components are reconstructed and enhanced to obtain the enhanced feature signal. Based on the enhanced feature signal, dimensionless transformation and unified feature space mapping are performed to obtain the mapped feature vector. Then, multi-level hierarchical integration is performed on the mapped feature vector to obtain the fused comprehensive feature representation. Based on the fused comprehensive feature representation, time-series evolution law modeling and dynamic prediction deviation analysis are performed to obtain an abnormal feature set; Based on the set of abnormal features, a multidimensional state trajectory is constructed. The evolution trend of the multidimensional state trajectory and the probability of failure risk are calculated using a long short-term memory network to obtain the failure prediction result of the intelligent device.

2. The intelligent device fault prediction method based on AI multimodal large model according to claim 1, characterized in that, The process of acquiring multi-source data streams during the operation of intelligent devices involves classifying and grouping these data streams and aligning them spatiotemporally with heterogeneous signals to obtain structured multi-source signal data, including: Acquire multi-source data streams from various sensors; Based on the multi-source data stream, physical characteristic identification is parsed and acquisition frequency values ​​are extracted to obtain the original signal sequence; Based on the original signal sequence, channel mapping based on preset physical characteristic identifiers is performed to obtain heterogeneous signal groups containing high-frequency and low-frequency signals. Based on the heterogeneous signal grouping, downsampling interpolation is performed for the high-frequency signal and forward hold timing alignment is performed for the low-frequency signal to obtain the aligned signal sequence; Based on the aligned signal sequence, time-domain distribution characteristics are calculated and structured model mapping is performed to obtain structured multi-source signal data.

3. The intelligent device fault prediction method based on AI multimodal large model according to claim 1, characterized in that, The process involves extracting sensor drift characteristic parameters and calculating deviation distribution based on the structured multi-source signal data to obtain the deviation distribution. Then, a recursive least squares algorithm is used to dynamically correct the signal deviation in real time based on this deviation distribution, resulting in a calibrated steady-state signal sequence, including: Based on the structured multi-source signal data, segmented statistical analysis is performed to obtain the signal amplitude within the current time window. The signal amplitude within the current time window is then compared with the pre-stored historical stable period reference amplitude to obtain the zero-point offset and sensitivity attenuation. Based on the zero-point offset and the sensitivity attenuation, frequency band mapping is performed to obtain the deviation distribution; Based on the deviation distribution, iterative calculations based on the recursive least squares algorithm are performed until the preset error tolerance range is reached to obtain dynamic compensation parameters; Based on the dynamic compensation parameters and the structured multi-source signal data, linear weighting and multiplication correction are performed to obtain the calibrated steady-state signal sequence.

4. The intelligent device fault prediction method based on AI multimodal large model according to claim 1, characterized in that, The step involves separating background noise and extracting weak fluctuation components based on the calibrated steady-state signal sequence, and then reconstructing and enhancing the weak fluctuation components to obtain an enhanced feature signal, including: Based on the calibrated steady-state signal sequence, the local signal-to-noise ratio is calculated, and signal segments with local signal-to-noise ratios lower than a preset weak feature screening threshold are selected to obtain a weak feature subset. Based on the weak feature subset, wavelet packet decomposition and soft threshold shrinkage denoising are performed to obtain a denoised feature vector sequence. Based on the denoised feature vector sequence, Butterworth low-pass filtering and residual Hilbert transform analysis are performed to obtain the early anomaly hidden signal; Based on the early anomaly hidden signal, sinusoidal superposition reconstruction and reference waveform superposition are performed to obtain the enhanced feature signal; The method of obtaining a weak feature subset further includes: For vibration or acoustic signature-type high-frequency non-stationary signals in the calibrated steady-state signal sequence, calculate the kurtosis or spectral kurtosis of the corresponding signal segment; Signal segments with kurtosis or spectral kurtosis values ​​higher than a preset kurtosis threshold are selected as part of the weak feature subset.

5. The intelligent device fault prediction method based on AI multimodal large model according to claim 1, characterized in that, The process involves performing dimensionless transformation and unified feature space mapping based on the enhanced feature signal to obtain mapped feature vectors, and then performing multi-level hierarchical integration on the mapped feature vectors to obtain a fused comprehensive feature representation, including: Based on the enhanced feature signal, statistical distribution moments are calculated and dimensionless operations are performed to obtain a normalized sequence. Based on the normalized sequence, principal component analysis projection and L1 norm sparsity constraint are performed to obtain the mapped feature vector. Based on the mapped feature vector, frequency band hierarchy is divided and time delay deviation is calculated to obtain the association weight of each level; Based on the association weights and the mapped feature vectors, a weighted concatenation and integration are performed to obtain the fused comprehensive feature representation.

6. The intelligent device fault prediction method based on AI multimodal large model according to claim 1, characterized in that, The process involves modeling the temporal evolution pattern and performing dynamic prediction deviation analysis based on the fused comprehensive feature representation to obtain an abnormal feature set, including: Based on the fused comprehensive feature representation, time-series evolution law learning and historical fluctuation variance analysis are performed to obtain the baseline prediction sequence and dynamic confidence interval; Based on the baseline prediction sequence and the fused comprehensive feature representation, the difference is calculated to obtain the residual sequence; Based on the residual sequence and the dynamic confidence interval, out-of-bounds point identification and likelihood calculation are performed to obtain potential outliers; Based on the potential anomalies, cluster center features are extracted to obtain an anomaly feature set.

7. The intelligent device fault prediction method based on AI multimodal large model according to claim 1, characterized in that, The process involves constructing a multidimensional state trajectory based on the set of abnormal features, and using a long short-term memory network to analyze the evolution trend and calculate the probability of failure risk on the multidimensional state trajectory to obtain the fault prediction result for the intelligent device, including: Based on the set of abnormal features, key parameters are extracted and multi-dimensional spatial mapping is performed to obtain state trajectory points; Based on the state trajectory points, a time-series evolution analysis is performed using a pre-trained long short-term memory network to obtain a dynamic trend curve; Based on the dynamic trend curve, the slope of the tangent is calculated to obtain the rate of change of the state trajectory points; Based on the rate of change of the state trajectory points, probabilistic inference is performed using a preset fault risk assessment model to obtain the fault prediction result of the intelligent device.

8. A fault prediction system for intelligent equipment based on an AI multimodal large model, characterized in that, include: The data processing module is used to acquire multi-source data streams during the operation of intelligent devices, classify and group the multi-source data streams and align them with heterogeneous signals in time and space to obtain structured multi-source signal data. The drift calibration module is used to extract sensor drift characteristic parameters and calculate deviation distribution based on the structured multi-source signal data, obtain the deviation distribution, and use the recursive least squares algorithm to perform real-time dynamic correction of signal deviation based on the deviation distribution to obtain the calibrated steady-state signal sequence. The signal enhancement module is used to separate background noise and extract weak fluctuation components based on the calibrated steady-state signal sequence, and to reconstruct and enhance the weak fluctuation components to obtain the enhanced feature signal. The feature fusion module is used to perform dimensionless transformation and unified feature space mapping based on the enhanced feature signal to obtain the mapped feature vector, and to perform multi-level hierarchical integration on the mapped feature vector to obtain the fused comprehensive feature representation. An anomaly detection module is used to perform time-series evolution law modeling and dynamic prediction deviation analysis based on the fused comprehensive feature representation to obtain an anomaly feature set; The risk prediction module is used to construct a multi-dimensional state trajectory based on the set of abnormal features, and to perform evolution trend analysis and fault risk probability calculation on the multi-dimensional state trajectory using a long short-term memory network to obtain the fault prediction result of the intelligent device.

Images